DETReg Incorporating Semi-Supervised Learning for Object Detection in the Advanced Driver-Assistance Systems

Abstract

In Advanced Driver-Assistance Systems (ADAS) and automatic driving, it is important to accurately recognize objects around the vehicle. DETReg is one of the unsupervised pre-training methods using Transformer, which is self-supervised by combining localization and categorization. DETReg performs self-supervised learning on unlabeled images. Then, it extracted a wide range of features from rich aspects of the data and gained the flexibility to adapt to many variations. Fine tuning then used the labeled dataset of the target task to fine tune the model to fit the specific dataset. This allowed DETReg to achieve higher accuracy in the object detection task. However, it is difficult to learn DETReg efficiently because of its slow learning time. In this paper, we propose a new pre-training method for object detection, called Semi-DETReg, that utilizes a few supervised labels during self-supervised learning. We incorporate semi-supervised learning into DETReg by using a portion of the supervised training data in the pre-training to improve efficiency. We demonstrate the effectiveness of our method by conducting experiments and comparing our method to a similarly trained DETReg.